Using Cure Models to Estimate Biological Cure Margaret R. Stedman, Ph.D. MPH Angela B. Mariotto, Ph.D. Data Modeling Branch Surveillance Research Program Division of Cancer Control and Population Sciences National Cancer Institute NAACCR Presentation June 5, 2012
Statistical Cure Fraction of cancer patients whose mortality rate equals the mortality rate of the cancer free population. Excess mortalityiszero is zero. The cured fraction is assumed to never experience a cancer death. Estimate of biological cure (population based measure) NOT personal cure probability of an individual dying of non cancer related causes.
Survival Estimates Colorectal Cancer, Regional Stage, 55 64 year olds All Cause 1 0.9 0.8 0.7 0.6 Su urvival Rate 0.5 0.4 0.3 0.2 0.1 0 0 5 10 15 20 25 30 Years from Diagnosis
Survival Estimates Colorectal Cancer, Regional Stage, 55 64 year olds 1 All Cause Cancer Specific 0.9 08 0.8 0.7 0.6 Survival Rate 0.5 0.4 0.3 0.2 0.1 0 0 5 10 15 20 25 30 Years from Diagnosis
Survival Estimates Colorectal Cancer, Regional Stage, 55 64 year olds All Cause Cancer Specific Cured Fraction 48% 1 0.9 08 0.8 0.7 0.6 Survival Rate 0.5 0.4 0.3 0.2 0.1 0 0 5 10 15 20 25 30 Years from Diagnosis
Survival Estimates Colorectal Cancer, Regional Stage, 55 64 year olds All Cause Cancer Specific Cured Fraction 48% 1 0.9 08 0.8 0.7 0.6 Survival Rate 0.5 0.4 0.3 0.2 CURED FRACTION 0.1 0 0 5 10 15 20 25 30 Years from Diagnosis
Importance of measuring Statistical Cure Patient survival is one of the most important questions in cancer research. Cure models give us more information about patient survival Predictproportion of patients cured from cancer. Predict time until cured Estimates survival time of patients not cured. Net survival estimate not influenced by other cause mortality
Cure is difficult to estimate Results can be unreliable S i b i df h Sometimes cure cannot be estimated from the data
Regional Stage Breast Cancer, 45-54 year olds. Lognormal Model Cure=35.0% Loglogistic Model Cure=39.6% Weibull Model Cure=47.4% Gompertz 4% Model Cure=46% Weibull Model Cure=47.4%
Criteria Follow up time > 2/3 median survival* % Coefficient of Variation (CV) < 10% Cure < last survival estimate *Yu et al., Statistics in Medicine, 2004 Rosner, Fundamentals of Biostatistics, 1995
Purpose To investigate criteria to assess the reliability of estimates from mixture cure models. To develop a method to easily summarize and evaluate if cure can be estimated. To identify cancer subgroups where a cure fraction can be estimated.
Example Study of Breast and Colorectal Cancer Study population SEER 9: San Francisco Oakland SMSA, 1973+','Connecticut 1973+','Detroit (Metropolitan) 1973+','Hawaii 1973+','Iowa 1973+','New Mexico 1973+','Seattle (Puget Sound) 1974+','Utah 1973+','Atlanta (Metropolitan) 1975+ Sl Selection criteria: i Years of Diagnosis: 1975 2007 Known Age: 45 74 Years Cases actively followed with malignant behavior Female Exclusions Death certificate and autopsy only diagnosis Secondary and later primaries Those alive without survival time Those with missing death certificate information
Example Study of Breast and Colorectal Cancer (CRC) Study population lti Site Age N Localized Regional Distant Unknown Breast 45 54 89365 56% 36% 5% 2% 55 64 95681 58% 33% 7% 2% 65 74 90039 62% 29% 7% 2% CRC 45 54 14832 37% 39% 21% 3% 55 64 28515 36% 39% 21% 3% 65 74 44243 38% 38% 20% 4%
Example Study of Breast and Colorectal Cancer Methods SEER*Stat software to obtain standard life tables of net survival (survival in the absence ofother other causes of death) Stratified by SEER historic stage, agegroup group Maximum follow up time 33 years. Begin at date of diagnosis (1975 2007) End at death or study cutoff (December 2008)
Example Study of Breast and Colorectal Cancer Methods 2 Estimates of Net Survival: Relative Survival (Ederer II method) Ratio of observed survival to survival of US general population Cause Specific Survival Uses death certificate information to determine if death is cancer related Censors other cause death.
Example Study of Breast and Colorectal Cancer Methods Parametric Mixture Cure Model: S(t) = C + (1 C)G(t; µ, σ) Net Survival Cured Fraction Uncured Fraction * Survival Function Tested 3 distributions for G(t): Distribution Survival FunctionG(t) Parameterization 1 Lognormal G(t;µ,σ) = 1 Φ[(ln{t} µ)/ σ] µ=µ σ=σ 2 Loglogistic G(t;λ,ρ) = [1 (λt) ρ ] 1 µ= ln λ σ=1/ρ 3 Weibull G(t;λ,ρ) = exp[1 (λt) ρ ] µ= ln λ σ=1/ρ
Example Study of Breast and Colorectal Cancer Methods Used Cansurv software to fit Mixture Cure Model Free software available at: http://surveillance.cancer.gov/cansurv/index.html
Results Statistics Cure estimate and confidence interval Mediansurvival timefor the uncured Last actuarial survival estimate and confidence interval cure last actuarial survival point Years of Follow-up median survival
Cause Specific Relative
Last Actuarial Survival [CI] Cure [CI] Cause Specific Relative
Last Actuarial Survival [CI] Cure [CI] Cause Specific Confidence Interval Relative?
Wide CI Models did not converge Weibull overestimates cure
Criteria Cure < survival %CV < 10% median time < 2/3 FU
Most reliable estimates Separation between causespecific and relative survival * younger age * regional, distant stage groups
Models did not converge Wide CIs C S Cause-Specific models overestimate cure
Criteria Cure < survival %CV < 10% median time < 2/3 FU
Most reliable estimates * All stages aged 45-54 * Relative survival
Cure Estimates Summary Stage Age Breast Cancer Colorectal Cancer L 45-54 66% - 87% 80% - 87% L 55-64 67% - 83% 63% - 79% L 65-74 73% 51% - 66% R 45-54 54 41% - 47% 55% - 59% R 55-64 35% - 40% 52% R 65-74 21% - 32% 48% - 54% D 45-54 8% - 12% 7% - 8% D 55-64 8% 6% D 65-74 6% 4% - 6% All 45-54 52% - 65% 54% - 56% All 55-64 33% - 55% 51% - 52% All 65-74 43% - 55% 43% - 47%
Site Breast cancer Conclusions subgroups Regional, distant stage Cause specific survival Colorectal cancer Relative survival Stage : Regional Localized stage and inconsistent estimates Increase in median survival relative to follow up time Differences in survival tail for Cause Specific and Relative. Distant stage and convergence problems. cure estimates are close to zero. Age: 45 54, 55 64 65 74 convergence problems, wide CIs Cure decreases slightly with age, but less effect with advanced stage disease.
Conclusions criteria Median time for uncured < 2/3 follow up localized disease with older age groups % CV < 10% Distant stage disease where estimates are close to zero Localized disease with older age groups. Cure < last survival estimate Wiblldi Weibull distribution ib ti All stages combined Regional colorectal cancer, cause specific estimates
Future Steps Additional cancer sites Adjust for diagnosis year in model Flexible models dl